Bacteria and other microorganisms navigate chemical gradients through a process called chemotaxis. At the microscopic level, they move by alternating straight runs with random tumbles, changing direction based on chemical signals in their environment. At the macroscopic level, the collective behavior of many such organisms can be described by smoother, averaged equations. A central challenge is understanding how these two levels of description connect, and specifically how to recover the mathematical rules governing when and how bacteria decide to tumble, given measurements of their behavior.
The paper focuses on estimating a mathematical object called the turning kernel, which encodes how bacteria reorient themselves in response to their surroundings. This kernel has two parts: one that controls basic movement and spreading, and another that drives directed migration up or down chemical gradients. The authors design a method based on minimizing a carefully constructed measure of error, and they prove that if this error measure is small, then the estimated kernel produces behavior that closely matches the true behavior, and importantly, this guarantee holds consistently whether you are looking at the microscopic or macroscopic scale. This scale-uniform stability is the key technical achievement, because it means the estimation method does not quietly break down as you zoom in or out.
On the practical side, the authors combine their theoretical framework with sparse inversion techniques, which are computational tools that favor simpler or more compact solutions when fitting data. This combination allows them to accurately recover turning kernels of various types, including smooth ones, irregular ones, and those that vary strongly across space, even when the measurements contain noise. The work essentially provides a principled bridge between microscopic bacterial motion models and their large-scale continuum approximations, with reliable guarantees on how well the underlying biological rules can be identified from data.